1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-dwconv/unipass-sse-mul32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/dwconv.h>
20 #include <xnnpack/intrinsics-polyfill.h>
21
22
xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul32(size_t channels,size_t output_width,const int8_t ** input,const void * weights,int8_t * output,size_t input_stride,size_t output_increment,size_t input_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__xop_mul32(
24 size_t channels,
25 size_t output_width,
26 const int8_t** input,
27 const void* weights,
28 int8_t* output,
29 size_t input_stride,
30 size_t output_increment,
31 size_t input_offset,
32 const int8_t* zero,
33 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(channels != 0);
36 assert(output_width != 0);
37
38 do {
39 const int8_t* i0 = input[0];
40 assert(i0 != NULL);
41 if XNN_UNPREDICTABLE(i0 != zero) {
42 i0 = (const int8_t*) ((uintptr_t) i0 + input_offset);
43 }
44 const int8_t* i1 = input[1];
45 assert(i1 != NULL);
46 if XNN_UNPREDICTABLE(i1 != zero) {
47 i1 = (const int8_t*) ((uintptr_t) i1 + input_offset);
48 }
49 const int8_t* i2 = input[2];
50 assert(i2 != NULL);
51 if XNN_UNPREDICTABLE(i2 != zero) {
52 i2 = (const int8_t*) ((uintptr_t) i2 + input_offset);
53 }
54 const int8_t* i3 = input[3];
55 assert(i3 != NULL);
56 if XNN_UNPREDICTABLE(i3 != zero) {
57 i3 = (const int8_t*) ((uintptr_t) i3 + input_offset);
58 }
59 const int8_t* i4 = input[4];
60 assert(i4 != NULL);
61 if XNN_UNPREDICTABLE(i4 != zero) {
62 i4 = (const int8_t*) ((uintptr_t) i4 + input_offset);
63 }
64 const int8_t* i5 = input[5];
65 assert(i5 != NULL);
66 if XNN_UNPREDICTABLE(i5 != zero) {
67 i5 = (const int8_t*) ((uintptr_t) i5 + input_offset);
68 }
69 const int8_t* i6 = input[6];
70 assert(i6 != NULL);
71 if XNN_UNPREDICTABLE(i6 != zero) {
72 i6 = (const int8_t*) ((uintptr_t) i6 + input_offset);
73 }
74 const int8_t* i7 = input[7];
75 assert(i7 != NULL);
76 if XNN_UNPREDICTABLE(i7 != zero) {
77 i7 = (const int8_t*) ((uintptr_t) i7 + input_offset);
78 }
79 const int8_t* i8 = input[8];
80 assert(i8 != NULL);
81 if XNN_UNPREDICTABLE(i8 != zero) {
82 i8 = (const int8_t*) ((uintptr_t) i8 + input_offset);
83 }
84 input = (const int8_t**) ((uintptr_t) input + input_stride);
85
86 size_t c = channels;
87 const void* w = weights;
88 for (; c >= 8; c -= 8) {
89 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
90 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((const int32_t*) w + 4));
91
92
93 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
94 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 0 * sizeof(int8_t))));
95 const __m128i vi0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0 + 4));
96 const __m128i vk0x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 4 * sizeof(int8_t))));
97 i0 += 8;
98
99 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
100 vacc4567 = _mm_macc_epi32(vi0x4567, vk0x4567, vacc4567);
101
102 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
103 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 8 * sizeof(int8_t))));
104 const __m128i vi1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1 + 4));
105 const __m128i vk1x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 12 * sizeof(int8_t))));
106 i1 += 8;
107
108 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
109 vacc4567 = _mm_macc_epi32(vi1x4567, vk1x4567, vacc4567);
110
111 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
112 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 16 * sizeof(int8_t))));
113 const __m128i vi2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2 + 4));
114 const __m128i vk2x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 20 * sizeof(int8_t))));
115 i2 += 8;
116
117 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
118 vacc4567 = _mm_macc_epi32(vi2x4567, vk2x4567, vacc4567);
119
120 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
121 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 24 * sizeof(int8_t))));
122 const __m128i vi3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3 + 4));
123 const __m128i vk3x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 28 * sizeof(int8_t))));
124 i3 += 8;
125
126 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
127 vacc4567 = _mm_macc_epi32(vi3x4567, vk3x4567, vacc4567);
128
129 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
130 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 32 * sizeof(int8_t))));
131 const __m128i vi4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4 + 4));
132 const __m128i vk4x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 36 * sizeof(int8_t))));
133 i4 += 8;
134
135 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
136 vacc4567 = _mm_macc_epi32(vi4x4567, vk4x4567, vacc4567);
137
138 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
139 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 40 * sizeof(int8_t))));
140 const __m128i vi5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5 + 4));
141 const __m128i vk5x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 44 * sizeof(int8_t))));
142 i5 += 8;
143
144 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
145 vacc4567 = _mm_macc_epi32(vi5x4567, vk5x4567, vacc4567);
146
147 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
148 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 48 * sizeof(int8_t))));
149 const __m128i vi6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6 + 4));
150 const __m128i vk6x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 52 * sizeof(int8_t))));
151 i6 += 8;
152
153 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
154 vacc4567 = _mm_macc_epi32(vi6x4567, vk6x4567, vacc4567);
155
156 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
157 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 56 * sizeof(int8_t))));
158 const __m128i vi7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7 + 4));
159 const __m128i vk7x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 60 * sizeof(int8_t))));
160 i7 += 8;
161
162 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
163 vacc4567 = _mm_macc_epi32(vi7x4567, vk7x4567, vacc4567);
164
165 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
166 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 64 * sizeof(int8_t))));
167 const __m128i vi8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8 + 4));
168 const __m128i vk8x4567 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 68 * sizeof(int8_t))));
169 i8 += 8;
170
171 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
172 vacc4567 = _mm_macc_epi32(vi8x4567, vk8x4567, vacc4567);
173
174 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t) + 72 * sizeof(int8_t));
175
176 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
177 __m128 vscaled4567 = _mm_cvtepi32_ps(vacc4567);
178
179 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
180 vscaled0123 = _mm_mul_ps(vscaled0123, vscale);
181 vscaled4567 = _mm_mul_ps(vscaled4567, vscale);
182
183 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
184 vscaled0123 = _mm_min_ps(vscaled0123, voutput_max_less_zero_point);
185 vscaled4567 = _mm_min_ps(vscaled4567, voutput_max_less_zero_point);
186
187 vacc0123 = _mm_cvtps_epi32(vscaled0123);
188 vacc4567 = _mm_cvtps_epi32(vscaled4567);
189
190 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
191 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
192
193 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
194 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
195 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
196
197 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
198 output += 8;
199 }
200 if XNN_UNLIKELY(c != 0) {
201 const int8_t* k = (const int8_t*) ((const int32_t*) w + 8);
202 do {
203 __m128i vacc0123 = _mm_loadu_si128((const __m128i*) w);
204
205 const __m128i vi0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i0));
206 const __m128i vk0x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(k));
207 i0 += 4;
208
209 vacc0123 = _mm_macc_epi32(vi0x0123, vk0x0123, vacc0123);
210 const __m128i vi1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i1));
211 const __m128i vk1x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 8)));
212 i1 += 4;
213
214 vacc0123 = _mm_macc_epi32(vi1x0123, vk1x0123, vacc0123);
215 const __m128i vi2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i2));
216 const __m128i vk2x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 16)));
217 i2 += 4;
218
219 vacc0123 = _mm_macc_epi32(vi2x0123, vk2x0123, vacc0123);
220 const __m128i vi3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i3));
221 const __m128i vk3x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 24)));
222 i3 += 4;
223
224 vacc0123 = _mm_macc_epi32(vi3x0123, vk3x0123, vacc0123);
225 const __m128i vi4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i4));
226 const __m128i vk4x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 32)));
227 i4 += 4;
228
229 vacc0123 = _mm_macc_epi32(vi4x0123, vk4x0123, vacc0123);
230 const __m128i vi5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i5));
231 const __m128i vk5x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 40)));
232 i5 += 4;
233
234 vacc0123 = _mm_macc_epi32(vi5x0123, vk5x0123, vacc0123);
235 const __m128i vi6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i6));
236 const __m128i vk6x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 48)));
237 i6 += 4;
238
239 vacc0123 = _mm_macc_epi32(vi6x0123, vk6x0123, vacc0123);
240 const __m128i vi7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i7));
241 const __m128i vk7x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 56)));
242 i7 += 4;
243
244 vacc0123 = _mm_macc_epi32(vi7x0123, vk7x0123, vacc0123);
245 const __m128i vi8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(i8));
246 const __m128i vk8x0123 = _mm_cvtepi8_epi32(_mm_loadu_si32((const void*) (k + 64)));
247 i8 += 4;
248
249 vacc0123 = _mm_macc_epi32(vi8x0123, vk8x0123, vacc0123);
250
251 k += 4;
252
253 __m128 vscaled0123 = _mm_cvtepi32_ps(vacc0123);
254 vscaled0123 = _mm_mul_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.scale));
255 vscaled0123 = _mm_min_ps(vscaled0123, _mm_load_ps(params->fp32_sse4.output_max_less_zero_point));
256 vacc0123 = _mm_cvtps_epi32(vscaled0123);
257
258 w = (const void*) ((const int32_t*) w + 4);
259
260 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
261 __m128i vout0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc0123), voutput_zero_point);
262
263 vout0123 = _mm_packs_epi16(vout0123, vout0123);
264 vout0123 = _mm_max_epi8(vout0123, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
265
266 if XNN_LIKELY(c >= 4) {
267 _mm_storeu_si32(output, vout0123);
268 output += 4;
269 c -= 4;
270 } else {
271 if (c & 2) {
272 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123, 0);
273 vout0123 = _mm_srli_epi32(vout0123, 16);
274 output += 2;
275 }
276 if (c & 1) {
277 *output = (int8_t) _mm_extract_epi8(vout0123, 0);
278 output += 1;
279 }
280 c = 0;
281 }
282 } while (c != 0);
283 }
284
285 output = (int8_t*) ((uintptr_t) output + output_increment);
286 } while (--output_width != 0);
287 }
288